Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
The AIR Professional FileFall 2019 Volume
Supporting quality data and decisions for higher education.
Association for Institutional Research© Copyright 2019, Association for Institutional Research
This volume of the AIR Professional File illustrates
creative approaches to data analysis in two familiar
areas of institutional research – class size and
enrollment projection.
In Perception Isn’t Everything: The Reality of Class
Size, Umbricht and Stange challenge traditional
measures of class size found in institutional fact
books, used by ranking publications, and computed
for the Common Data Set. What if, instead, we
evaluated class size through the lens of the student
experience? The authors explain how to compute a
student-centric metric of class size and offer advice
for the effective use of this more sophisticated and
nuanced measure.
Who among us hasn’t been pressed to find a magic
formula to accurately forecasts student enrollment?
In Enrollment Projection Using Markov Chains, Gandy,
Crosby, Luna, Kasper and Kendrick show how an
underutilized forecasting tool can be enlisted to
better understand and predict student flow, even in
subgroups of students. With extensive description
of the methodology and presentation of results, the
authors provide a blueprint for replication at your
institution.
Do you have a creative approach to data analysis?
Your colleagues are waiting for you to share it
through the AIR Professional File!
Sincerely,
Sharron Ronco
Editors
Sharron Ronco
Coordinating Editor
Marquette University (retired)
Stephan C. Cooley
Managing Editor
Association for Institutional Research
ISSN 2155-7535
LETTER FROM THE EDITOR
Perception Isn’t Everything: The Reality of Class Size
Article 146 Authors: Mark Umbricht and Kevin Stange
Enrollment Projection Using Markov Chains: Detecting Leaky Pipes and the Bulge in the Boa
Article 147 Authors: Rex Gandy, Lynne Crosby, Andrew Luna, Daniel Kasper, and Sherry Kendrick
Thank You
About the AIR Professional File
More About AIR
Table of Contents
4
21
39
39
40
4Fall 2019 Volume
Mark Umbricht and Kevin Stange
About the Authors
The authors are with the University of Michigan.
Kevin Stange is also with the National Bureau of
Economic Research.
Acknowledgments
We would like to thank Paul Courant, Ben Koester,
Steve Lonn, Tim McCay, and numerous participants
in seminars at the University of Michigan for helpful
comments. This research was conducted as part
of the University of Michigan Institutional Learning
Analytics Committee. The authors are solely
responsible for the conclusions and findings.
Abstract
Every term, institutions of higher education must make decisions about the class size for each class they offer,
which can have implications for student outcomes, satisfaction, and cost. These decisions must be made
within the current higher education landscape of tightening budgets and calls for increased productivity.
Beyond institution decision making, prospective students and their families may use class size as one factor
in deciding whether an institution might be a good fit for them. The current measure of class size found in
university fact books, and subsequently sent to numerous ranking groups such as U.S. News & World Report
(hereafter U.S. News), is an inadequate gauge of the student experience in the classroom, as measured
by the percent of time students spend in classes of varying sizes. The current measure does not weight
for enrollment, credits, or multiple components of a class, which results in a misleading representation of
the student experience of class size. This paper will discuss these issues in depth, explain how class size
varies across institutions, and offer recommendations on how to reweight class size in the Common Data
Set to accurately describe it from the student’s perspective. Institutions could use this new metric to better
understand class size, and subsequently to understand the student experience and cost of a class, while
prospective students and their families could use the metric to gain a clearer picture of the class sizes they
are likely to experience on campus.
Keywords: class size, productivity, student outcomes
Perception Isn’t Everything: The Reality of Class Size
The AIR Professional File, Fall 2019
Article 146
https://doi.org/10.34315/apf1462019
© Copyright 2019, Association for Institutional Research
5Fall 2019 Volume
INTRODUCTIONOrganizing class delivery is a key operational
decision for institutions of higher education. Each
term these institutions must decide how many
students will be taught within each section given
the classes they offer, the faculty and instructors
they have available to teach, and the confines of
physical spaces they have on campus. Within these
constraints, institutions must decide how to deliver
classes. Consider a popular class taken by nearly
every first-year student: Should this class be taught
as one large lecture by a famous professor, many
small sections taught by graduate students, or a
combination of the two? Should small classes target
freshmen who are acclimating to college or seniors
as they specialize in their field?
Institutions, particularly public institutions, make
these decisions within the current context of
increased accountability and decreased resources.
Traditional wisdom argues that smaller classes
increase engagement, facilitate student-faculty
interactions, and improve student success. The
opportunity to learn from prominent scholars in the
field is also considered a strength of undergraduate
education at research universities. However, smaller
classes and senior faculty are costlier and their use
comes at the expense of other ways of enriching
and supporting the undergraduate student
experience. As Courant and Turner (forthcoming)
argue, institutions have an interest in providing
curriculum efficiently, meaning they must strike a
balance between quality, costs, and tuition revenue.
If an institution or department has an influx of
students, it must decide whether it will increase
the size of its faculty or the class size of its courses.
Therefore, decisions about class size have first-order
influence on student success and institutional costs.
From the student perspective, class size could be
influential in the college choice process, with some
students seeking intimate class settings with small
class sizes, and others preferring to blend in to a
large classroom. Students and their families rely
on institutional websites and rankings, such as
Princeton Review or U.S. News, and other publicly
available data for information about class size. These
data are typically drawn from the Common Data Set
(CDS), which is a collaborative effort among data
providers and publishers to improve the quality and
accuracy of information provided to prospective
students, and to reduce the reporting burden on
data providers (CDS Initiative, 2018). While it is
helpful to have a measure that can be reported
across multiple campuses, the class size metric
used by the CDS is measured at the classroom level
rather than at the student level. This difference in
measurement leads to a disconnect between the
metric and the phenomena it is trying to describe, as
described in the following example.
Imagine a high school student researching her
nearby public, research university as a prospective
student. She sees that, according to U.S. News
in 2018, only 17% of classes have more than 50
students, and 57% of classes have fewer than 20
students. The student thinks, What luck! She thinks
she can attend a high-quality research institution
while spending most of her time in small classes.
After graduation from that college, the same
student looks back and sees that she spent more
than 41% of her time in classes with more than
50 students, and only 20% of her time in classes
with fewer than 20 students. These differences in
the perception versus reality are not exaggerated,
but rather are many students’ average experience.
This paper will show that the measure of class size
calculated for the CDS, and subsequently used by
many other sources, does not provide an accurate
approximation of the true class size experienced
6Fall 2019 Volume
by students at the University of Michigan (U-M), a
large, research university in the Midwest. By the
term “student experience,” we mean the percent
of time or credits spent in classes of varying sizes.
The results of this case study could be replicated at
any institution, with varying degrees of departure
from the true student experience depending on the
institution type and size. Specifically, this paper will
argue for a new class size metric to be used in the
CDS and will address the following questions:
Framing Questions
1| How does the standard definition of class size
vary from the student experience?
2| How does a student-centric version of class size
vary across an institution?
3| How can institutional researchers practically
recalculate class size to better approximate
the student experience without significantly
increasing the burden of data providers?
Importance of the Topic and Literature Review
This topic is important for students, institutions,
and the field of institutional research. From the
student perspective, students and those assisting
in their decisions need accurate and meaningful
information to make the best decision about which
college to attend. A small number of studies have
shown that class size is an important factor for
students as they select an institution (Drewes &
Michael, 2006; Espinoza, Bradshaw, & Hausman,
2002). This makes sense since lower class size is
perceived to be linked to gains in student outcomes.
Literature in the secondary setting is clear that lower
class size is associated with gains across multiple
areas, including test scores, noncognitive skills,
college enrollment, and other outcomes (Angrist &
Lavy, 1999; Chetty et al., 2010; Dee & West, 2011;
Dynarski, Hyman, & Schanzenbach, 2013; Hoxby,
2000; Krueger, 1999). However, in higher education
the relationship between class size and outcomes
is not well established, with studies finding either
negligible association (Bettinger, Doss, Loeb, Rogers,
& Taylor, 2017; Lande, Wright, & Bartholomew,
2016; Stange & Umbricht, 2018; Wright, Bergom,
& Lande, 2015) or a negative relationship between
class size and outcomes (Bettinger & Long, 2018; De
Giorgi, Pellizzari, & Woolston, 2012; Kokkelenberg,
Dillon, & Christy, 2008). Institutions that gain a more
accurate and more nuanced version of class size
from the student experience perspective could
aid prospective students in their decision-making
process.
Class size is also important to institutions
for planning purposes. Courant and Turner
(forthcoming) argue that institutions must strike a
balance between quality, costs, and tuition revenue.
In recent years, institutions have been asked to
cut back and do more with fewer resources, which
would imply that increasing class size would be an
appropriate strategy. In fact, class size is one of
the most important drivers of instructional costs
(Hemelt, Stange, Furquim, Simon, & Sawyer, 2018).
However, lower class size is perceived to lead to
better student outcomes and is subsequently tied to
rankings such as those at U.S. News. This common
perception pulls institutions to keep class size lower,
putting institutions in a situation where a logical
solution is to hire cheaper instructors, such as
noncontingent faculty. The ultimate decision on how
to strike this balance is not traditionally made at the
institution level, but rather at the department level.
Cross and Goldenberg (2009) found that the number
of noncontingent faculty at elite research institutions
rose significantly in the 1990s, which was due to
7Fall 2019 Volume
micro- (department-)level decisions. Departments
(or colleges) that are particularly concerned with
the quality (or perceived quality) of small class sizes
would find it difficult to adequately assess how
much time their students spend in classes of a given
size with the current metric, which is at the class
level. A new student experience version of class
size would allow departments to compare the class
size experience across multiple majors or between
departments, which could assist in balancing
the class size constraints for long-term planning.
Having an accurate understanding of imbalances by
class size across colleges, departments, or majors
could help institutions pinpoint areas that need
improvement. In addition, institutions could examine
whether access to smaller classes is inequitable
across certain student groups, such as among
minority, first-generation, or first-year students.
We will argue that the current definition of class size
used in the CDS Initiative (2018) is insufficient for
both internal planning and external consumption.
As institutional researchers, it is our job to provide
meaningful and accurate information to both
internal and external parties. While the traditional
measure of class size may be accurate, this paper
describes ways in which we could provide data that
are more meaningful. Institutional researchers and
higher education professionals have an obligation
to update this metric to reflect the actual student
experience.
DEFINING CLASS SIZEBased on the conventions of the CDS,
undergraduate class size is calculated based on
the number of classes with a given class size range.
Classes are divided into sections and subsections.
A class section is an organized class that is offered
by credit, is identified by discipline and number,
meets at one or more stated times in a classroom or
setting, and is not a subsection such as a laboratory
or discussion section. A class subsection is a part of
a class that is supplementary and meets separate
from the lecture, such as laboratory, recitation, and
discussions sections. In calculations of class size,
we count only the sections of a class and discard
the subsections. The CDS conventions consider
any section or subsection with at least one degree-
seeking undergraduate student enrolled for credit
to be an undergraduate class section, but exclude
distance learning, noncredit, and specialized one-
on-one classes such as dissertation or thesis, music
instruction, one-to-one readings, independent study,
internships, and so on. If multiple classes are cross-
listed, then the set of classes are listed only once to
avoid duplication (CDS Initiative, 2018). This means
that we count stand-alone classes, defined as having
only one component, once per section in the class
section portion.
For classes with multiple components, such as a
lecture section combined with a lab or discussion
section, we count each lecture section once in the
class section portion while we count each associated
lab or discussion section once in the class
subsection table. In traditional class size metrics,
the CDS counts only the class section portion of the
class while the CDS discards the subsection from the
calculation. This metric is relatively easy to compute
and is comparable across campuses, but it may not
be representative of the student experience.
We define “student experience class size” as the
percent of time spent by a student in classes of
various sizes, using credits as a proxy for time.1
Calculations for this metric will be discussed later in
1. We assume that each credit associated with a class is approximately 50 minutes of class time. While this pattern is not universal across U-M, the calculation is easy to make and should be readily accessible in any institution’s data warehouse. A more complicated, but 100% accurate approach, would be to use the day and time location to derive the true number of minutes spent in each section. These data might not be accessible and would require substantial coding to calculate. We tested both approaches and our simplified approach did not meaningfully differ.
8Fall 2019 Volume
this paper. Figure 1 shows the difference between
the CDS method and our new student experience
method of computing class size. Sources that are
often used as references for prospective students
and institutions, such as U.S. News and institutional
websites, draw data from the CDS. The CDS metric
describes the share of classes in a given range
rather than the share of time spent in classes of
varying sizes. According to the U.S. News and the
U-M websites, 84% of classes at U-M have fewer
than 50 students, and 57% have fewer than 20
students. However, using our student experience
class size metric, only 19% of a student’s classroom
time is spent in classes with fewer than 20 students
and nearly 30% of their time is spent in classes with
at least 100 other students. Why do these metrics
differ so drastically?
There are three primary reasons driving these
differences. First, the traditional measure for class
size is not weighted by the number of students
enrolled. A 500-student section and a 5-student
section both count as one class, even though
many more students experience the larger section.
Second, classes are not weighted by the number of
credits associated with the class. A class worth five
credits counts for the same as a class worth one
credit, even though students likely spend five times
as much time in the first class. Finally, the traditional
measure does not incorporate subsections. It is
typical for large lecture classes to have multiple
components, such as a large lecture of 200 that
meets for 2 hours per week and 10 associated small
discussion groups of 20 students each that meet
for 1 hour per week. Students spend 67% of their
classroom time in a large lecture and 33% of their
time in a small discussion, but the traditional metric
Figure 1. Class Size by Various Sources
9Fall 2019 Volume
counts only the lecture portion. This means the
200-person lecture counts as one class, ignoring the
subsections. Our new student experience class size
metric accounts for these three factors, as will be
explained in detail in the methods section. This new
metric provides a more accurate representation of
the student experience within the classroom.
Given the immense difference between the
traditional class size measure and our student
experience version, it is only natural to question why
institutions have not moved to a different calculation
of class size. To be clear, it is not the authors’ belief
that institutions are purposely trying to push an
inaccurate measure of class size. The traditional
class size measure does hold value in describing the
number of classes available to students of various
class sizes. U-M students can choose from many
small classes and could theoretically construct a set
of classes to minimize the amount of time spent in
large classes. In reality, though, several forces could
make it difficult for institutions to switch to a student
experience version of class size.
First, there are serious consequences in rankings
and optics for many institutions, particularly larger
ones. U.S. News currently provides points to
institutions based on the share of small sections (19
and fewer students), partial points for the share of
medium sections (20–49 students), and no points for
large sections (50 or more students).2 Universities
that use larger class sizes to teach a large number
of students will see their rankings negatively
impacted if classes were weighted by the number
of students taking the class. The shift in class size
would also create poor optics for prospective
students and may impact whether they choose one
institution over another. The same would be true
for a student experience measurement of instructor
type, where larger research institutions are much
more likely to use graduate students as instructors.
In addition, having non-tenure track instructors
primarily responsible for teaching large classes, and
therefore many students, could have bad optics
for institutions. Therefore, there is a disincentive
for a single college, or a small group of colleges, to
recalculate their class size based on the student
experience. The exception would be institutions
that uniformly have very small classes, such as
small liberal arts colleges, that would see little or no
change in their calculation of class size.
A second difficulty is measuring class size from the
student perspective. Leaders of the CDS Initiative
already consider the measurement of class size to
be the second-most difficult part of the CDS, with
only calculations of financial aid deemed more
difficult (Bernstein, Sauermelch, Morse, & Lebo,
2015). At U-M, measures required to recalculate
class size to the student perspective are readily
available and clean, and require little manipulation
to combine. Institutions vary significantly in their
data capacity and availability of staff to adjust the
CDS measures. Given these challenges, the authors
of this study still believe that shifting to a student
experience version of class size would provide many
benefits, including an accurate representation of
the amount of time students spend in classes of
varying sizes and with various instructor types. This
shift would be beneficial for institutions for planning
purposes as well as for prospective students as
they weigh various institutions during the selection
process.
2. This recently changed from a system that provided points for small courses and penalties for large courses in an effort to minimize gaming of the system (Supiano, 2018). However, there is no evidence provided to back up this claim. Regardless of whether the new or old system is used, having a larger number of small classes is rewarded, and the metric is at the course level, which does not appropriately describe the student experience.
10Fall 2019 Volume
METHODSThe purpose of this study is to create a student
experience version of class size. We drew data
from the U-M data warehouse, specifically from the
Learning Analytics Data Architecture (LARC) and
College Resource Analysis System (CRAS).3 LARC is a
flattened, research-friendly version of the raw data
warehouse that houses data about students, their
background, their progress, and their coursework.
CRAS is a data warehouse system that houses data
about classes and the instructors that teach them.
The sample included first-time freshman students in
cohorts between 2001 and 2012, examining classes
taken within 4 years of entry. Freshman cohorts are
between approximately 5,500 and 6,500 students
during this period. Individual study classes and one-
on-one classes were removed from the sample, as
were classes with no CRAS information, including
subjects such as medicine, dentistry, armed forces,
study abroad, and classes through the Committee
on Institutional Cooperation program. The final
sample included 70,426 first-time students and
3,398,320 class sections taken in these students’ first
4 years of study, between 2001 and 2016.
Calculating Class Size
As previously noted, we made three adjustments to
the traditional measure of class size: (1) weighting
for number of students, (2) weighting for credits
associated with the section, and (3) incorporating
subsections. Table 1 provides an example of how
3. LARC is unique to U-M, although some institutions have developed a similar research-friendly database. CRAS is also unique to U-M because there are some calculations made with institution-specific formulas. However, these databases comprise data that are regularly available (although not always clean) at all institutions, such as student and class information, classes taken, and the number of students in a given class.
Table 1. Example of Distribution of Class Credits in Our Framework
Number of Students Class 1 Class 2 Class 3 Class 4
Total Credits
% of Total
2–9
10–19
20–29 1 (Lab) 1 (Discussion) 2 17%
30–39
40–49 3 (Lecture) 3 25%
50–99 2 (Lecture) 2 17%
100–199 2 (Lecture) 2 17%
200+ 3 (Lecture) 3 25%
Total Credits
3 4 3 2 12 100%
11Fall 2019 Volume
we accounted for credits and subsections in our
student experience framework.
This example shows a hypothetical student’s
coursework for one term. This student took four
classes in this term for a total of 12 credits. Classes
1 and 2 had multiple components, with a lecture
and either a discussion or a lab. We divided credits
for these classes among the section and subsection
based on the amount of time or credits associated
with each component. Then we put each component
into the appropriate class size bin, shown in the left
column. Classes 3 and 4 were stand-alone classes
that contained only a lecture component, so there
was no need to divide their credits. Once we had
distributed all the credits for each class to the
appropriate class size bin, we totaled credits across
each class size row. In doing so, we accounted for
credits and the multicomponent nature of classes,
fixing two of the issues with the standard definition
of class size. The third piece relates to weighting
class size by the number of students in the class. In
this framework, we theoretically create a table like
this for each student and each term the student
attends. We then summed across every student
and term. Since the level of observation is a class
enrollment, we naturally weight by the number
of students in the class because there will be 50
observations if there are 50 students in a class, or
5 observations for a class with 5 students. A basic
assumption made by this framework is that one
credit is equal to approximately 1 hour of class
time. While there are a small number of classes that
violate this assumption, we do not believe it would
impact our overall results in a meaningful way. It is
also important to note that enrollments for cross-
listed classes were combined into the home class.
At U-M, if there are multiple cross-listed sections,
one is considered the home class and the rest are
considered away classes. This means that if there
are three cross-listed sections with 12, 15, and 18
students, our data would show one class with 45
students.
Once we calculated the percent of time (using
credits as a proxy) that a student earned in various
class sizes across his first four years, we calculated
percentiles for each enrollment group across the
entire university by college, by major, and by year
in school.4 College and major are determined by
the last college or major associated with a student.
If a student graduated with a bachelor’s degree, we
used his graduating college or major. If a student
departed prior to completing his degree, we used
the last known college and major associated with
him.5 Rather than showing just the median or mean,
we chose to use five percentiles (10th, 25th, 50th,
75th, and 90th) to show the distribution of student
experiences. Finally, we mapped these percentiles
into figures that show the range of student
experiences for a given class size.
CASE STUDYThis section will examine how class size varies
across U-M. The figures in this section represent
the distribution of time that students spend in
classrooms of varying sizes. Class size was grouped
into eight bins of varying sizes to create a smooth
figure and to mimic the traditional measure of
class size. Figure 2 shows the distribution of class
size across the entire university. The black line
represents the median student, the dark gray
shaded area represents the 25th to 75th percentile,
and the light gray shaded area represents the 10th
4. Results show the 10th, 25th, 50th, 75th, and 90th percentile for each enrollment group. The term “college” refers to an academic college, such as engineering, liberal arts and science, or education.
5. We considered removing these students but decided that doing so could introduce some selection bias. For example, students that drop out may select larger courses; if we remove them, we may distort the student experience.
12Fall 2019 Volume
to 90th percentiles.
To interpret this figure, consider the enrollment
group of 200+ students. The black line indicates that
the median U-M student spent about 15% of her
time in classes with 200 or more students. Moving
above the black line to the edge of the dark gray
area, we see that about a quarter of students spent
about 23%–24% of their time in classes with 200 or
more students. The outer edge of the light gray area
indicates that 10% of students at U-M spent more
than 30% of their time in these very large classes. If
we move to the 20–29 enrollment group, we see that
the median student spent about 20% of her time
in classes with 20–29 students. There is a very wide
range of experiences (25%) between the 10th and
90th percentiles, indicating that students may spend
vastly different amounts of time in classes with
20–29 students. The spread is only about 10% wide
for enrollment groups between 30 and 49 students,
indicating there is less variability in the percent of
time spent in medium-size classes. Overall, it is clear
that students’ time is more heavily weighted in both
large (50+ students) and small (10–29 students)
classes. This means that students spend their time
in classes of varying sizes, but classes at U-M tend to
favor high or low enrollments on average. This also
implies that very large classes likely have a smaller
enrollment component tied to them, such as a
discussion or lab section.
At a university the size of U-M, with an
undergraduate population of almost 30,000, it
is natural to assume that student experiences
may vary greatly across the institution, such as by
college, major, or year in school. Figure 3 shows
the distribution of class size in the College of
Engineering, which we chose because it is the
Figure 2. Percent of Time Spent by Size Across the University of Michigan
13Fall 2019 Volume
second-largest college on campus and differs
significantly from the trends for the median student.
Once again, the black line and gray shaded areas
represent the percentiles for students in a given
college. We added the gray dotted line here to show
the median U-M student from Figure 2, allowing
us to examine how the college differs from the
overall pattern at the university. In the College of
Engineering students tend to take coursework with
very different class sizes compared to the median
U-M student. In particular, engineering students
take fewer small classes (10–29 students) and
very large classes (200+ students), and instead
take more classes in the middle range (30–199
students). Part of the difference could be attributed
to deliberate planning by the College of Engineering
and part could be related to the size of classrooms
in the engineering buildings. Classroom caps,
and subsequently enrollment caps, for classes
in engineering tend to lie in the middle of the
enrollment group distribution.
While not shown, we created figures for every
college on campus. The trends of these figures show
that class size differs significantly across colleges.
The College of Literature, Science, and the Arts
(LSA) has very similar trends to the median U-M
student, in part because it is the largest college on
campus. Given that more than 60% of students
are in LSA, this college drives much of the median
class size. As one would expect, smaller and more
narrowly focused colleges such as the College of
Architecture and Urban Planning, the College of
Art and Design, and the College of Music, Theatre,
and Dance had significantly higher levels of small
classes (2–29 students) and fewer very large classes
(200+ students). The College of Public Policy tends
to mimic the trends of LSA, in part because students
Figure 3. Percent of Time Spent by Enrollment Group in the College of Engineering
14Fall 2019 Volume
spend their first two years in LSA before declaring
their major. The College of Business had very high
levels of medium-size classes (40–99 students)
because many of its core classes have enrollments
of 40 to 80 students.
Class size also varies in systematic ways by
major, even within a college. Comparing the
major of Arts and Ideas in the Humanities, a
small, multidisciplinary major, to the major of
Biopsychology, Cognition, and Neuroscience (BCN),
a large, premed major, yielded large differences in
class size. The median Arts and Ideas major spent
about 50% of her time in classes with 20 or fewer
students, which is twice the 25% of the median LSA
student. Students in the BCN major, on the other
hand, took a much larger share of large lectures,
rising out of the 90th percentile for LSA students.
They spent nearly 33% of their time in classes with
more than 200 students, compared to only 20% of
the median LSA student’s time. While we observed
systematic differences between majors, we also
found that there were some majors that had very
similar class size structures.
A final way to observe how class size varies across
an institution is by comparing class size by academic
level. Rather than aggregating across a college
or major, we aggregated by a student’s year in
school (e.g., freshman, sophomore, junior, senior).
Students in their first year typically fulfill their
general education requirements, which tend to be
classes taught to many students at once. By their
junior or senior year, students tend to take many
classes within their major of increasing depth and
specialization, characterized by smaller class sizes.
Splitting the data in this manner did yield interesting
differences over time. As shown by the dotted line in
Figure 4. Class Size Within the College of Engineering by Academic Year
15Fall 2019 Volume
Figure 4, students in the College of Engineering took
similar coursework to LSA students in their first year
compared to the median U-M student from Figure
2, taking mainly classes consisting of large lecture
and small discussion groups. This is likely because
students are fulfilling their general education
requirements. By their third year, the solid black line
indicates the median engineering student deviates
from the pattern at the college level and takes a
much higher proportion of classes with 50–100
other students; these classes comprise classes in
the students’ major.
Overall, we created figures for every major, college,
academic level, and student across campus to
examine how class size varied across U-M. The
median student within some majors had class size
distributions that were very similar to the median
student in their college and the university, but many
majors differed significantly from the general trends.
Similarly, there were some students within majors
that varied significantly from the median student in
their major. These figures show that class size can
vary significantly across an institution. A department,
college, or institution can use this information to gain
a more nuanced understanding of the experiences
of their students. For example, if students that
tend to take only large lectures or small discussion
classes perform worse, then an institution could
adjust its advising to promote students to take
classes of varying sizes. Similarly, an institution could
identify whether certain student groups, such as
first-generation students, may benefit from a more
intimate classroom environment where they receive
more attention from instructors.
HOW SHOULD WE MEASURE CLASS SIZE?This paper has shown that the traditional measure
of class size is not sufficient if it is meant to provide
information about a student’s actual experience
in the classroom. Previous research has shown
that increasing class size has a mixed but generally
negative impact on student learning and satisfaction.
Accurately measuring class size is also important
to institutions because it impacts productivity. For
example, increasing or decreasing the class size of
introductory calculus, a class that most students on
campus take, can have vast implications for the cost
of the class. If an institution wanted to recalculate
class size with the student experience at the core,
what would it look like?
We will first consider two simple adjustments:
weighting for credits and weighting for students.
Table 2 shows the distribution of class size given
four different calculations and Figure 5 provides
a visual representation of the table. The first
calculation is the traditional measure of class size,
with no adjustments. This means there is one
observation per lecture. The second column weights
the traditional measure by number of credits. A
class of four credits is now worth twice as much as
a class of two credits. This slightly shifts the class
size distribution down because large class sections
have corresponding subsections. Consider a class
worth three credits, two earned in a lecture and one
earned in a lab. Since the lab, or subsection, is not
counted in the traditional measure, one of the three
credits is discarded, deflating the value of a large
class. The third column accounts for subsections
and credits, appropriately distributing all the credits
associated with each class. A class worth three
credits including a lecture and a lab, as described
above, is now fully included in the metric for class
16Fall 2019 Volume
size (Column 3). This lowers class size considerably
because we are not removing the subsection but
instead including it in our calculation. Subsections
tend to be smaller, which drives the distribution
down.
The fourth column shows class size distribution if
we weighted only for the number of students in a
class. Using this calculation, a large lecture of 500
students is counted 500 times, while a small class
of 5 students is counted only 5 times. As expected,
this dramatically shifts the distribution of class size
upward, nearly flipping the distribution of class size
from the traditional measure. Column 5 accounts
for weighting by credits and students, and includes
subsections simultaneously. This gets closer to
the perfect measure of the student experience
described earlier in this paper and vastly improves
the understanding of class size for institutions
and potential students. Figure 5 shows a visual
representation of each strategy for calculating
class size. We can see that the traditional measure
(nonstudent experience), weighting for credits, and
weighting for subsections are very similar, with
slight increasing low enrollments (fewer than 30)
and decreasing other enrollments (30–200+) after
accounting for subsections. Once we weight for
students there is an immediate shift, nearly flipping
the distribution. However, this shift goes too far;
accounting for credits and subsections provides a
balance between the three strategies.
It is important to note that our calculation of class
size used student-level micro data to account for
these changes, in part due to other related research.
However, this calculation may be burdensome for
institutions. A simpler way to achieve the same
results would be to take the class-level data used for
Table 2. Class Size by Three Different Calculations
Class Size
Traditional Measures
(1)
Weight for Credits
(2)
Account for Subsections
(3)
Weight for Students
(4)
Weight for All Changes
(5)
2–9 13.9% 14.1% 11.4% 2.0% 2.5%
10–19 32.1% 32.5% 31.1% 13.0% 15.4%
20–29 23.0% 23.6% 34.0% 14.2% 21.9%
30–39 8.2% 8.5% 9.4% 6.8% 9.3%
40–49 4.4% 4.3% 3.2% 4.6% 4.4%
50–99 10.5% 10.2% 6.4% 18.2% 15.8%
100–199 4.9% 4.5% 2.9% 17.2% 13.8%
200+ 2.9% 2.4% 1.7% 24.0% 16.8%
17Fall 2019 Volume
the CDS and simply weight each class section and
subsection by the number of students and credits.
This simplifies the process and limits the resources
required to pull and process the data.
LIMITATIONS OF NEW CLASS SIZE MEASUREMENTWhile this new class size metric may provide a
clearer picture of the student experience at an
institution, it does not come without limitations
and challenges. The first challenge is the
technical barriers of calculating the new class size
measurement. For example, institutions may house
the required variables in different data systems, but
replacing or adding a new metric could be difficult to
implement and could increase the reporting burden
on institutions. Most of what is required to adjust
to this new metric is already required for the CDS
version of class size. Institutions must calculate the
number of sections and subsections of a given class
size. This means they must know the exact class size
for every section and subsection. The only missing
piece is to allocate time between the section and
the subsection. At U-M, a field for distributed hours
is contained in the data warehouse, but that may
not be the case for all institutions, some of which
may not even use the Carnegie credit system. For
these institutions, a calculation could be made to
allocate credits based on the amount of classroom
time for each section and subsection. For example, a
three-credit class that meets for 2 hours in a lecture
section and 1 hour in a discussion section could split
the class into two credits for the lecture and one
Figure 5. Class Meeting Size Distribution by Five Different Methods
18Fall 2019 Volume
credit for the discussion. This would require some
up-front work to create these calculations based on
day/time information, but is not unmanageable.
The second set of challenges relate to the nuance
that is inherent in the new metric. Presenting
students with a single metric for an entire university
is simple and easy to explain but leaves out a lot of
nuance. If an institution were to recreate the figures
in this paper, it would introduce some challenges
for interpretation. As shown in Figure 3, there are
stark differences in class size between engineering
and LSA students; the same differences can be
shown between majors or year in school. The large
number combinations and comparisons by college,
major, and year in school present challenges when
trying to show or explain the data to prospective
or current students. While a single, institution-wide
metric leaves out nuance, it is a vast improvement
over the current metric used by institutions and
rankings. If an institution wanted to provide more
nuance, it could create an interactive dashboard
for prospective and current students to view the
nuanced version of class size presented in the
figures of this paper. Students could select a college
or major from a list to see what the distribution
of class size looks like for students in that major.
Institutions could pare down these figures to
show only the median for simplicity or to provide
a detailed explanation of how the percentiles
work when students first use the new dashboard.
However, with a nuanced view an institution could
also show students that class size may be lower for a
given major, or how class size may lower as students
progress toward their degrees.
DISCUSSION AND FUTURE WORKAs institutional researchers it is imperative for us to
provide data that are both accurate and meaningful.
This paper argues that the traditional measure of
class size is not a meaningful representation of
what students experience in the classroom. The
traditional measure of class size illustrates only the
proportion of classes that are small, not the amount
of time that students spend in small classes. This
is problematic for prospective students, who could
use meaningful class size data to determine where
they want to attend college. It is also problematic
for institutions, which may not understand the
extent of large or small classroom experiences on
their campuses. While limited, previous research in
higher education suggests that class size matters for
student outcomes and satisfaction in classes.
This paper suggests that the measurement of class
size could be altered by weighting for the number
of students and credits associated with the section,
and accounting for subsections. We suggest that
institutional researchers consider revamping their
class size metric to reflect the student experience
in the classroom more accurately. Nearly all the
required components (number of students in
each class and section/subsection, and number of
classes) for this calculation are used by the current
CDS metric. We believe that distributed credits, the
potential missing component, is likely captured and
readily available at many institutions, which would
make this adjustment relatively easy. While this could
require an investment of time on behalf of some
institutions, we believe the potential benefits will
outweigh the investment. At a minimum, we suggest
that institutional research professionals consider
reweighting their current metric of class size by the
number of students in each lecture section. This new
19Fall 2019 Volume
metric provides a more accurate description of class
size from the student experience.
Future work on this topic could yield more
improvements in the description of the student
experience at institutions of higher education. First,
the range of class size also varies significantly across
colleges and departments, indicating that a simple
institution-wide metric masks important differences
for students who plan to major in different fields.
Institutions could take this idea one step further
to create an interactive dashboard that allows
prospective and current students the opportunity
to see the range of class size experiences for majors
in which they have interest. A second improvement
institutions could make would be to pair data
about class size and instructor type (e.g., IPEDS
instructor type). By combining the percent of time
spent in varying class sizes and varying instructor
types, students will gain a clearer picture of what
their classroom experience would be at a particular
institution.
REFERENCESAngrist, J., & Lavy, V. (1999). Using Maimonides
Rule to estimate the effect of class size on
scholastic achievement. Quarterly Journal of
Economics, 114(2), 553–575. https://www.jstor.org/
stable/2587016?seq=1#page_scan_tab_contents
Bernstein, S., Sauermelch, S., Morse, R., & Lebo,
C. (2015). Fundamentals and best practices for
reporting Common Data Set (CDS) data. Presented
at the Association for Institutional Research Annual
Conference, Denver, CO, May 27.
Bettinger, E., Doss, C., Loeb, S., Rogers, A., &
Taylor, E. (2017). The effects of class size in online
college courses: Experimental evidence. Economics
of Education Review, 58, 68–85. https://doi.
org/10.1016/j.econedurev.2017.03.006
Bettinger, E. P., & Long, B. T. (2018). Mass instruction
or higher learning? The impact of college class size
on student retention and graduation. Education
Finance and Policy, 13(1), 97–118. https://doi.
org/10.1162/edfp_a_00221
Chetty, R., Friedman, J. N., Hilger, N., Saez, E.,
Schanzenbach, D. W., & Yagan, D. (2011). How
does your kindergarten classroom affect your
earnings? Evidence from Project STAR. Quarterly
Journal of Economics, 126(4), 1593–1660. https://doi.
org/10.1093/qje/qjr041
Common Data Set (CDS) Initiative (2018). 2018–
2019 Common Data Set Initiative. http://www.
commondataset.org/
Courant, P. N., & Turner, S. (forthcoming). Faculty
deployment in research universities. In Hoxby, C. M.
& Stange, K. (Eds.) Productivity in higher education.
Chicago: University of Chicago Press.
Cross, J. G., & Goldberg, E. N. (2009). Off-track profs:
Nontenured teachers in higher education. Cambridge,
MA: MIT Press.
De Giorgi, G., Pellizzari, M., & Woolston, W. G. (2012).
Class size and class heterogeneity. Journal of the
European Economic Association, 10(4), 795–830.
https://onlinelibrary.wiley.com/doi/abs/10.1111/
j.1542-4774.2012.01073.x
20Fall 2019 Volume
Dee, T. S., & West, M. R. (2011). The non-cognitive
returns to class size. Educational Evaluation
and Policy Analysis, 33(1), 23–46. https://doi.
org/10.3102/0162373710392370
Drewes, T., & Michael, C. (2006). How do students
choose a university?: An analysis of applications to
universities in Ontario, Canada. Research in Higher
Education, 47(7), 781–800. https://doi.org/10.1007/
s11162-006-9015-6
Dynarski, S., Hyman, J., & Schanzenbach, D. (2013).
Experimental evidence on the effect of childhood
investments on postsecondary attainment and
degree completion. Journal of Policy Analysis
and Management, 32(4), 692–717. https://doi.
org/10.1002/pam.21715
Espinoza, S., Bradshaw, G., & Hausman, C. (2002).
The importance of college choice factors from the
perspective of high school counselors. College and
University, 77(4), 19–23.
Hemelt, S. W., Stange, K. M., Furquim, F., Simon, A.,
& Sawyer, J. E. (2018). Why is math cheaper than
English? Understanding cost differences in higher
education. Working Paper 25314, National Bureau
of Economic Research, Cambridge, MA. https://doi.
org/10.3386/w25314
Hoxby, C. M. (2000). The effects of class size
on student achievement: New evidence from
population variation. Quarterly Journal of
Economics, 115(4), 1239–1285. https://doi.
org/10.1162/003355300555060
Kokkelenberg, E. C., Dillon, M., & Christy, S. M.
(2008). The effects of class size on student grades
at a public university. Economics of Education
Review, 27(2), 221–233. https://doi.org/10.1016/j.
econedurev.2006.09.011
Krueger, A. B. (1999). Experimental estimates
of education production functions. Quarterly
Journal of Economics, 114(2), 497–532. https://doi.
org/10.1162/003355399556052
Lande, E., Wright, M., & Bartholomew, T. (2016,
October). Impact of smaller class sizes in U-M
mathematics courses. Unpublished internal report,
Center for Research on Learning and Teaching,
University of Michigan, Ann Arbor.
Stange, K., & M. Umbricht, 2018. Undergraduate
class size reduction: Effects and costs. Unpublished
working paper, University of Michigan, Ann Arbor.
Supiano, B. (2018, March 21). Are small classes
best? It’s complicated. Chronicle of Higher Education.
https://www.chronicle.com/article/Are-Small-Classes-
Best-It-s/242878
Wright, M., Bergom, I., & Lande, E. (2015,
September). Impact of smaller class sizes in U-M
Spanish and German courses. Internal report,
Center for Research on Learning and Teaching,
University of Michigan, Ann Arbor. Executive
summary at http://crlt.umich.edu/sites/default/files/
SmallClassExecSum.pdf
21Fall 2019 Volume
Rex Gandy, Lynne Crosby, Andrew Luna, Daniel Kasper, and Sherry Kendrick
About the Authors
The authors are with Austin Peay State University.
Abstract
While Markov chains are widely used in business and industry, they are used within higher education
only sporadically. Furthermore, when used to predict enrollment progression, most of these models use
student level as the classification variable. This study uses grouped earned student credit hours to track
the movement of students from one academic term to the other to better identify where students enter or
leave the institution. Results from this study indicate a high level of predictability from one year to the next.
In addition, the use of the credit hour flow matrix can aid administrators in identifying trends and anomalies
within the institution’s enrollment management process.
Keywords: Markov chains, enrollment projections, enrollment management, enrollment trends, enrollment
Enrollment Projection Using Markov Chains: Detecting Leaky Pipes and the Bulge in the Boa
The AIR Professional File, Fall 2019
Article 147
https://doi.org/10.34315/apf1472019
© Copyright 2019, Association for Institutional Research
22Fall 2019 Volume
INTRODUCTIONThe current challenges facing higher education
administrators create myriad reasons to find
a crystal ball of sorts to effectively forecast
enrollments, predict how many current students will
stay at the institution, forecast new students, and
adequately estimate revenues. These challenges
have become only more pressing in recent years.
More than 20 years ago, when public college and
university revenues were ample, administrators
were not readily concerned about the future of
college enrollments or student persistence. State
appropriations were healthy and usually made up
more than half of an institution’s revenue source.
Moreover, with lower tuition more students could
afford to obtain a degree without going into
significant financial debt (Coomes, 2000).
The costs to run higher education have skyrocketed,
however, causing today’s institutions to seek scarce
resources within an ever-diminishing financial
pool. As states tackle other pressing issues such
as infrastructure, entitlements, and prisons, the
amount they give to higher education naturally
wanes. Decreased state revenue, therefore,
compels institutions to increase tuition to make
up the difference. According to Seltzer (2017),
for every $1,000 cut from per student state and
local appropriations, the average student can be
expected to pay $257 more per year in tuition and
fees. He further notes that this rate is rising.
In addition to decreases in state revenues, higher
education administrators are under increasing
pressure to be accountable to federal and state
governments as well as to regional and discipline-
based accreditors. This accountability is increasingly
seen in tougher reporting standards, outcomes-
based funding formulae, and mandated student
achievement thresholds.
The closest resource to a crystal ball available to
administrators is a set of mathematical prediction
tools. These prediction tools range from simple
formulae contained in spreadsheets to much more
complicated regression, autoregressive integrated
moving average (ARIMA), and econometric time
series models.
According to Day (1997), current predictive tools
that are statistically based rely on the institution’s
ability to access and manipulate large datasets
and individual student-record data. While more-
complicated statistical models incorporate variables
such as tuition cost, high school graduate numbers,
economic factors, and labor-market demand,
other models look more specifically at institutional
indicators such as high school grade point averages
of entering freshmen, as well as the retention,
progression, and graduation rates of students.
One such model, the Markov chain, has been
relatively underutilized as an enrollment projection
tool in higher education. When used properly,
however, it can aid institutions in determining
progression of students. Specifically, Markov chains
are unique from more-traditional ARIMA and
regression prediction tools in that the following is
true:
1| Markov chains can give accurate enrollment
predictions with only the previous year’s data.
These predictions can be helpful when large
longitudinal databases are not available.
2| They can generate predictions on segments of
a group of students rather than on the entire
population. Other models often require the use
of the entire population.
23Fall 2019 Volume
3| The almost intuitive nature of the Markov
chain lends well to changes in student flow
characteristics that often cannot be explained
by a complex statistical formula.
Moreover, Markov chains might be particularly
helpful in determining progression of students
during benchmark years when enrollments vary
significantly due to state mandates and policies,
or due to institutional changes in admission
standards. The purpose of this study is to show how
a Southeastern, masters-level (Larger Programs)
public institution utilized the unique properties of
this model to create a tool to better understand
credit hour flow and student persistence.
Enrollment Management’s Problem with Leaky Pipes and the Bulge in the Boa
While enrollment management has clearly evolved
since the inception of the field of enrollment
management in the 1970s, some fundamental
processes have essentially stayed the same.
Institutions have always wanted to attract the right
students who fit well within the institution’s role,
scope, and mission. Once matriculated into the
institution, there is also a strong desire for students
to adequately progress through their program
and graduate within a reasonable amount of
time (Hossler, 1984). As enrollment management
developed through time, however, administrators
became increasingly aware that college-age students
were more difficult to enroll, higher tuition was
causing some students to forgo their degree,
and institutional loyalty was waning as students
transferred to similar or different institutions.
Furthermore, institutions have seen an increasing
number of students who are not fully prepared for
the rigors of college work, putting greater enrollment
strain on institutions (Johnson, 2000).
After more than 40 years of enrollment management
within higher education, it is not surprising that
metaphorical associations have entered the lexicon
of the profession as administrators try to better
understand and predict student matriculation,
persistence, and graduation. For instance, Ewell
(1985), referred to students progressing and
moving throughout their program as student flow,
while Clagett (1991) discussed following the flow
of student cohorts through to graduation. Luna
(1999) used the concept of student flow to explain
the various pathways by which the institution may
retain students, and Torraco and Hamilton (2013)
discussed the student flow of selected groups of
minority students. Furthermore, many software
companies have exploited the student flow
metaphor to describe use of data to identify areas
where leakage is present in student flow pipelines. It
is easy, then, to see how the management of student
retention can be associated with a pipeline and how
administrators are busy trying to plug the leaks.
Markov chains are uniquely suited to identifying
these leaks because they can model student flow
as a set of transitions between several states, much
like a set of pipes with various inflows, outflows,
and interconnections. In addition to using the
model to project enrollments, it is also possible to
observe from year to year where students enter the
absorbed state (i.e., do not return to the institution).
Leakage within the student credit hour (SCH) flow
pipeline occurs when students withdraw or stop out
due to reasons that are academic, nonacademic, or
both. If the model can isolate where the major leaks
occur, the institution can identify causes and work
to retain and maintain the flow of students within
the pipeline. These leaks in the student flow pipeline
can be detected and monitored from term to term
so that the institution can develop strategies to
maintain a healthier flow.
24Fall 2019 Volume
Another colorful bit of jargon among enrollment
management professionals is the idea of
bulging enrollments. For example, Fallows and
Ganeshananthan (2004) use the term “bulging
of enrollments” to describe a significantly larger
share of students needing financial aid or when,
due to rising tuition costs, students bulge into
less-expensive 2-year colleges. Herron (1988) uses
the term “bulge in the boa” to define instances of
oversupply in student populations quickly entering
the student flow pipeline, much as a boa constrictor
swallows a large meal. Liljegren and Saks (2017)
added that these bulges can significantly affect
higher education and its future. These bulges occur
when large groups of students suddenly enter
higher education, putting a strain on the student
flow pipeline. As the bulge dissipates, its effects
may remain, and it may redefine student flow for
the future. With Markov chain models, institutions
can monitor these bulges in the system so that they
can address issues such as course offerings and
instructor availability.
Markov Chains and Higher Education
A Markov chain is a type of projection model
created by Russian mathematician Andrey Markov
around 1906. It uses a stochastic (random) process
to describe a sequence of events in which the
probability of each event depends only on the state
attained in the previous event.
The Markov chain is a stochastic rather than a
deterministic model. Unlike a deterministic process
where the output of the model is fully determined by
the parameter values and by sets of previous states
of these values, a stochastic process possesses
inherent randomness: the same set of parameter
values and initial conditions can lead to different
outputs.
Take, for example, the scenario of an individual
returning home from work. In a deterministic
process, there is only one route (Route A) from
work to home, and the amount of time to get home
depends only on the variable speed of the driver. In
a stochastic process, the individual will have multiple
routes (Routes A, B, and C) from which to choose,
and each of the routes intersects the other routes
at various points. The randomness of the process
occurs when the individual combines routes to go
home, if she makes the choices at each intersection
randomly. For example, the driver may take Route A
part of the time, followed by Route C, then Route B,
and back to Route A again, or take some completely
different path. There are many random possibilities
the individual may take to get home, leading to a
variety of possible driving times.
Markov chains utilize transition matrices that
represent the probabilities of transitioning from
each possible state to each other possible state.
These states can be absorbing or nonabsorbing:
nonabsorbing states allow future transitions to other
states while absorbing states do not.
Markov chains have been widely and successfully
used in business applications, from predicting sales
and stock prices to personnel planning and running
machines. Markov chains also have been used in
higher education, albeit with much less frequency.
In most studies where Markov chains were used in
enrollment management, the various transitional
states were categorized either by student
classification or by other simpler dichotomous
measures. Given the strength of the Markovian
stochastic process in generating student flow
probabilities using data only from the previous
year, the process of classifying students into other
kinds of states could be appealing. Such states
25Fall 2019 Volume
could include SCHs, student debt, and (on a more
systemwide level) the transitioning from one
institution or program to another. The possibilities
are diverse.
One of the first to use Markov chains in determining
enrollment projections was Oliver (1968) when
he compared Markov chains to the much more
established use (at that time) of grade progression
ratios to predict enrollments at the University of
California. According to Oliver’s study, enrollment
forecasting made a prediction on the basis of
historical information on past enrollment and
admission trends. In determining a stochastic
process, Oliver demonstrated that the fraction of
students who leave one grade level (class status) i
and progress to class status j is a fraction pij; that
progress could also be time dependent. These
fractions pij can also be interpreted as random
transition probabilities. He determined that the
process allowed for contributions in one grade level
that were identified by their origins, such as prior
grade level, returning to the same grade level, and
new admissions (Oliver, 1968).
According to Hopkins and Massy (1981), the use of
Markov chains allows the researcher to observe the
flow of students from one classification level (i.e.,
freshman, sophomore, junior, senior) to the next
class level. The chain also incorporates students who
stay at the same class level from one year to the
next. Therefore, the Markov chain for class level, as
studied by Hopkins and Massy, can be described as
follows:
1| The number of students in class level i who
progress to class level j
2| The number of students in class level i who stay
in the same level
3| The number of students who leave the
institution (drop out, stop out, or graduate)
Similarly, Borden and Dalphin (1998) used Markov
chains to develop a 1-year enrollment transition
matrix to track how students of each class level
progressed. The authors found that unique Markov
chain models were valuable in measuring student
progression without having to rely on 6-year
graduation rate models, which could be ineffective
due to the large time lags. Specifically, the model
was built around a transition matrix where student
flow was tracked from one year to the next, and the
rates of transition from four nonabsorption states
(i.e., freshman to sophomore) were placed into a
matrix that was separate from the two absorption
states (i.e., drop out, graduate).
Using the percentages in the two matrices, those
students who continue in nonabsorption states were
processed through the matrix using the established
rates of transition until, asymptotically, all students
reach the final absorption state.
Additionally, Borden and Dalphin (1998) developed
discrete Markov chain processes to simulate
the effect of changes in student body profile on
graduation rates. In these models, the authors
incorporated credit-load and grade performance
categories. Their results indicated that, while
there was a strong association between grade
performance and persistence, it took very large
changes in levels of student performance to impact
retention and graduation rates modestly.
In a more narrowly focused study, Gagne (2015)
used Markov chains to predict how English
Language Institute (ELI) students progressed
through science, technology, engineering, and
math (STEM) programs. Specifically, the model
26Fall 2019 Volume
created transitional (nonabsorbing) states based
on classification level and three absorbing states to
include those students who left the institution, those
who graduated from a STEM program, or those who
graduated from a non-STEM program. Findings from
their study indicated that the ELI students tended to
progress at a higher rate than non-ELI students in
STEM programs, and that ELI students who repeated
the freshman year were more likely to repeat again
than they were to transition to the sophomore year.
Correspondingly, a study by Pierre and Silver (2016)
used Markov chain models to determine the length
of time it took students to graduate from their
institutions. As with previous studies, students
were divided into nonabsorbing transitional states
(i.e., freshman, sophomore, junior, senior) and
absorbing states (i.e., graduate, nonreturning). Using
the Markovian property, the future probability of
transitioning from one state to another depended
only on the present state of the process and was not
influenced by its history. The study found that it took
5.9 years for a freshman to graduate and 4.5 years
for a sophomore to graduate from the institution.
Brezavšček, Bach, and Baggia (2017) successfully
used Markov chain models to investigate the pattern
of students’ enrollment and academic performance
at a Slovenian institution of higher education. The
model contained five transient or nonabsorbing
states and two absorbing states. The authors used
student records for a total of eight consecutive
academic seasons, and estimated the students’
progression toward the next stage of the program.
From those transition percentages they were able
to obtain progression, graduation, and withdrawal
probabilities.
As mentioned earlier, most Markov chain models
involving enrollment management and prediction
use student classification to create the various
states of the model. Using student classification in
model specification, however, could create states
that are overly broad in nature since, at most
semester-based colleges and universities, student
classification varies by 30 hours.
Ewell (1985), who also used Markov chains to
predict college enrollments, noted two limitations
of the models. First, because the estimation of the
probabilities rests on historical data, Markov chains
may be sensitive to when the data were collected.
This could be especially true with significant
enrollment gains or declines from one year to
the next. Second, according to Ewell, different
subpopulations may behave in different ways, thus
necessitating the need to disaggregate into smaller
groupings.
However, the Markov chain’s attributes may allow
a unique ability to detect the leaks and bulges.
Because this type of projection model uses the
stochastic process to describe a sequence of events
in which the probability of each event depends only
on the state attained in the previous event, changes
to student flow are immediate and are not subject to
potentially skewed results of the past. In short, the
limitations mentioned by Ewell (1985) can be utilized
when building the student flow matrices to detect
significant shifts in enrollment and to determine
which groups of students are leaving the institution
at a higher rate.
METHODOLOGYThe current study used Markov chains to predict
Fall enrollment at a Southeastern, masters-level
(Larger Programs) public institution based on
annual Fall semester enrollment for degree-seeking
27Fall 2019 Volume
undergraduates. The process involved obtaining
data from the institution’s student information
system and separating students into groupings
based on their cumulative SCHs earned. Student
flow was measured from Fall of year i to Fall of year
i+1 based on whether students stayed within their
credit hour category, moved into another credit
hour category, or did not enroll at the institution.
These student flow changes for each category were
then summed and applied to year i+2 as a prediction
of enrollment.
Within the model, at a given point in time each
student has a particular state, and each student
is treated as having a particular probability of
transitioning to each other state or staying within
the same state. Most of these states are based on
the number of SCH the student has accumulated
(i.e., the SCH category). Because the SCH category
of a student was determined by the number of
cumulative SCHs a student earned, most of the
credit hour flow scenarios included students
advancing to a higher credit hour category or
students withdrawing or graduating. While it is rare
for a student to move from a particular credit hour
category to a lower category, it can happen through
the transfer process when, after the student has
enrolled, the current institution does not accept
certain SCHs from the former institution.
The characteristic that makes this model a Markov
chain is the fact that a given student’s transition
probabilities between states are assumed to depend
only on that student’s current state and not on any
of the student’s previous states. This is a simplifying
assumption that allows all students within a given
state to be treated similarly regardless of their
histories. Otherwise, the model would become much
more complicated and difficult to apply.
The main parameters of the model are estimates
of these transition probabilities. These transition
probabilities are estimated by calculating the
fractions of students that transitioned from each
state to each other state relative to the number
of students initially in that state in past years’
enrollment data. The other parameters of the model
are the fractions of new incoming students by credit
hour category. The total number of new incoming
students is assumed to be fixed, thus the estimated
number of incoming students by credit hour
category follows from these fractions.
The model process is recursive in that predictions
for Fall X are produced from the enrollment data
from Fall X-2 and Fall X-1 and the subsequent flow
rates from Fall X-2 to Fall X-1.
We can now describe the basic assumptions that we
used to construct the predictive models:
1| Each model models flow from one year to the
next and is named accordingly. For example, Fall
2013 to Fall 2014 is known as the 13_14 Model
and is based on the starting data for Fall 2013
and the new student data from Fall 2014.
2| As the model is applied, the output headcount by
SCH level for the (i+1)th year becomes the input
headcount for the next iteration of the model.
3| When the model is applied to a future year, the
total number of new students is assumed to be
constant and the same as the number of new
students for the (i+1)th year. The distribution of
new students by SCH level is also assumed to be
constant.
4| When the model is applied to a future year, it
is assumed that the fractional student loss and
fractional student continuation ratios are fixed
by SCH level.
28Fall 2019 Volume
5| When the model is applied to a future year it is
assumed that the fractional flow from SCH level
to SCH level is the same as for the year used to
construct the model.
The model divides the undergraduates into 24
6-SCH groupings. This method uses historic
ratios of SCH student subsets gathered from the
student information system to predict future Fall
headcounts.
The 6-SCH groupings used in this model are
individually less broad than the more familiar
student classification levels. However, it is possible
to aggregate the 6-SCH bins into a version of these
student levels, which we define as
• Freshmen: ≤30 SCH
• Sophomore: >30 SCH and ≤60 SCH
• Junior: >60 SCH and ≤90 SCH
• Senior: >90 SCH
Note that these classification-level definitions do
not exactly match the institution’s definitions. In
using SCH groupings, the enrollment pipeline may
be much more finely observed and enrollment
patterns among students may be more precisely
distinguished. While it is the goal of this study
to develop a model to predict the coming Fall
enrollment once the previous Fall enrollment is
known, the model will not address enrollment by
major, academic department, or college.
MODEL DESCRIPTIONThe student information system parsed out students
into the various SCH categories based on the
predetermined groupings. These students were
then tracked during the following Fall semester to
determine student flow percentages. Within this
study, student flow states are defined as:
1| students in credit hour group j who stayed
within that group,
2| students in credit hour group j who moved to a
different credit hour group,
3| students in other credit hour groups who
moved to group j, and
4| students who were no longer enrolled at the
institution.
Within this model, the following terms and symbols
are used:
1| n is the number of SCH levels in the model (n =
24 for the 6-SCH groupings).
2| hij is the ith Fall semester headcount for the jth
SCH level.
3| Hi is the total undergraduate headcount for the
ith semester.
4| lij is the number of the hij subset students not
enrolled the next Fall semester.
5| Li is the total number of undergraduates
enrolled in the ith Fall semester that are not
enrolled in the (i+1)th Fall semester.
6| cij = hij – lij is number of continuing students in
the jth SCH level.
7| Ci is the total number of undergraduates that
enrolled in the ith Fall semester that are also
enrolled in the (i+1)th Fall semester.
29Fall 2019 Volume
8| dijk is the number of the continuing cij subset
students that move from SCH level j to SCH level
k from the ith Fall to the (i+1)th Fall.
9| wij is the number of the Ci subset students that
flow from all other levels into level j.
10| oij is the number of the cij subset students that
flow out of level j into all other levels.
11| s(i+1)j is the number of the new incoming students
for the (i+1)th Fall semester where j is the SCH
level.
12| N(i+1) is the total number of incoming new
undergraduate students for the (i+1)th
semester.
With this terminology in place, the previously stated
assumptions of the models can now be described
algebraically:
1| When applying a model to a future period
from Fall (i+1) to Fall (i+2), the total number of
incoming students is assumed to be the same
as it was for the period used to build the model,
so it is assumed to have the value Ni+1. The
fraction of new students by SCH level for that
upcoming year is also assumed to be the same
as it was in the period used to train the models,
so each is assumed to be s(i+1)j ⁄ Ni+1. Therefore,
the estimated number of new students for a
particular SCH level in that future year can be
obtained by multiplying the value of this fraction
by the estimated total number of students in
the current year. That is, the estimate for the
number of new students in the future year for
that particular SCH level is given by
s(i+1)j ⁄ Ni+1 × Ni+1 = s(i+1)j.
2| The fractional loss and fractional continuation
ratios are also assumed to be fixed by SCH level.
In other words, for a future year these ratios
are assumed to be lij ⁄ hij and cij ⁄ hij, the same as
they were in the year used to build the model.
Therefore, for the upcoming future period from
Fall (i+1) to Fall (i+2), the estimated number of
lost and continuing students for the jth SCH
level are obtained by multiplying these ratios by
the number of students h(i+1)j in that SCH level in
the current Fall (i+1). This multiplication is lij ⁄ hij
× h(i+1)j to estimate lost students in the jth SCH
level and cij ⁄ hij × h(i+1)j to estimate continuing
students in the jth SCH level.
3| Finally, the fractional flow from a particular SCH
level to another SCH level is assumed to be
fixed. In other words, for a future year these
ratios are assumed to be dijk ⁄ cij, the same as
they were in the year used to build the model.
Therefore, for the upcoming future period from
Fall (i+1) to Fall (i+2), the estimated number
of students transitioning from SCH level j to
SCH level k is given by the value of this ratio
dijk ⁄ cij multiplied by the estimated number of
continuing students in the jth SCH level.
The processes described above can be applied
iteratively to obtain estimates for years even farther
into the future by using the estimated values from
one iteration as inputs into the next iteration.
Using the terms and formulae, we created a
spreadsheet matrix (Table 1) that includes the
various credit hour classifications as well as the
nonabsorbed transient student states and the
absorbed state of no longer enrolled.
30Fall 2019 Volume
From this SCH flow structure, we can observe
the relationships of credit hour flow between
and among the various states, including flow into
nonabsorbing states (staying or moving into another
credit hour state) or into absorbing states (not
enrolling at the institution). The relationships among
the variables are as follows:
1| cij = ∑ nk=1 dijk represents those current students
who were in SCH level j who stayed at the
institution.
2| oij = ∑ nk=1k≠j
dijk represents those current students
who were in SCH level j who moved to all other
SCH levels.
3| wik = ∑ nj=1j≠k
dijk represents those current students
who were in SCH levels other than k who moved
to SCH level k.
4| Hi = ∑ nj=1 hij represents semester headcount at
Fall semester i.
5| Li = ∑ nj=1 lij represents those students at Fall
semester i who did not reenroll.
6| Ci = ∑ nj=1 cij represents those students at Fall
semester i who did reenroll.
The following relationship,
∑ ∑n
k=1
n
j=1wik = oij
shows two equivalent ways of expressing the
collection of students who remain at the institution
and move from any SCH level to a different SCH
level during the year. Conservation of student flow is
obtained only when students from level j stay in SCH
Table 1. Basic Structure Matrix of the Markov Chain ModelSC
H L
evel
s
Fall
i Hea
dcou
nt
Lost
Cont
inui
ngMovement from SCH Level
Stat
ic
Infl
ow
Out
flow
Fall
i+1,
New
Stu
dent
s
Fall
i+1
Hea
dcou
nt
1 2 . . n
1 hi1 li1 ci1
Mov
emen
t to
SCH
Lev
el 1 di11 di11 wi1 oi1 s(i+1)1 h(i+1)1
2 hi2 li2 ci2 2 di12 di22 di22 wi2 oi2 s(i+1)2 h(i+1)2
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
n hin lin cin n di1n . . . dinn dinn win oin s(i+1)n h(i+1)n
Note: This table shows the basic structure matrix of the headcount SCH flow associated with the Markov chain model that connects
the undergraduate headcount in the ith Fall to the headcount in the (i+1)th Fall.
31Fall 2019 Volume
level j or move to other SCH levels, or when students
from other SCH levels move into SCH level j.
Given these relationships, the number of
undergraduates by level in the second Fall semester
can be calculated using the following formula:
h(i+1)j = hij – ij – oij + wij + s(i+1)j
This is the number of total transient students in
one of the SCH levels after 1 year who were not
absorbed by withdrawing or graduating. Therefore,
the total number of students in the (i+1)th Fall
semester is simply given by
Hi+1 = Hi – Li – Ni+1
since the inflow and outflow terms cancel upon
summation.
RESULTSThe model used actual data from a Southeastern,
masters-level (Large Programs) public institution
for Fall 2010 through Fall 2017. The enrollments for
these 8 years are displayed in Table 2.
In developing the Markov chain matrix for each year,
the total number of students within each category
were noted and tracked to the following year. Within
this matrix, one can observe the various student
states by each category to determine who is moving
into transitional (nonabsorbing) states and who is
graduating or not returning. These more-granular
data within the matrix offer clues as to when
students may be leaving the institution and where
there are potential bulges in the system coming
from new or transfer students.
Table 2. Annual Enrollment Data, Fall 2010–Fall 2017
Fall iFall i
Headcount Lost Continuing NewFall (i+1)
Headcount
Fall 2010 9,652 3,773 5,879 3,957 9,836
Fall 2011 9,836 4,082 5,754 3,721 9,475
Fall 2012 9,475 3,965 5,510 3,761 9,271
Fall 2013 9,271 3,843 5,428 3,574 9,002
Fall 2014 9,002 3,685 5,317 3,598 8,915
Fall 2015 8,915 3,792 5,123 3,993 9,116
Fall 2016 9,116 3,945 5,171 3,919 9,090
Fall 2017 9,090 not known not known not known not known
32Fall 2019 Volume
28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1
SCH Level Groupings
(>162)
(157–162)
(151–156)
(145–150)
(139–144)
(133–138)
(127–132)
(121–126)
(115–120)
(109–114)
(103–108)
(97–102)
(91–96)
(85–90)
(79–84)
(73–78)
(67–72)
(61–66)
(55–60)
(49–54)
(43–48)
(37–42)
(31–36)
(25–30)
(19–24)
(13–18)
(7–12)
(0–6) SCH Level Definition
257
49 61 96
113
135
154
179
237
271
323
341
400
347
347
346
369
446
399
291
316
332
386
501
322
245
264
1589 HC1 (Fall 2016)
176
25 45 78 79 89
116
131
171
188
230
204
248
165
121
84 85
105
114
92 94
108
121
147
141
110
81
597 Lost
81 24 16 18 34 46 38 48 66 83 93
137
152
182
226
262
284
341
285
199
222
224
265
354
181
135
183
992 Continuing
4 1 4 2 0 1 1 3 5 3 5 8 11 9 13 18 16 37 31 19 16 13 17 11 1 1 0 0
GradA16
4 1 2 0 0 1 1 2 5 3 1 6 6 5 12 12 13 27 21 17 13 8 13 9 1 0 0 0
GradA16E
124
21 32 56 48 56 90 98
135
159
188
147
175
100
34 14 1 0 0 0 0 0 0 0 0 0 0 0
GradB16
2 1 0 0 1 3 2 3 3 2 5 1 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0GradB16E
Movement to SCH Level Number
28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1
1 3 15 60
258
408
158
60 14 1
Movem
ent from SCH
Level Num
ber
1 3 5 21 85 42 10 6 5 2
4 13 45 35 18 9 5 3
1 4 16 46 58 31 13 7 4
3 5 38
115
103
57 19 11 5
1 1 6 24 89 82 31 19 5 1 6
1 1 10 19 70 59 33 17 9 7
3 8 29 69 43 39 17 11 8
1 1 7 31 56 49 27 13 10 9
5 9 71 84 58 26 21 9 1 10
4 20 53
123
79 28 23 9 11
4 18 46 85 79 25 13 12 12
1 1 2 5 5 45 79 64 34 18 7 13
1 8 28 62 84 17 17 8 14
1 8 18 60 52 28 10 2 15
2 1 16 38 49 23 13 5 16
1 2 4 16 27 44 19 16 5 17
12 25 18 17 10 8 18
2 1 4 22 23 11 11 6 19
1 4 10 21 8 12 7 1 20
1 6 10 14 6 6 3 21
4 8 10 4 4 4 22
18 6 9 6 5 23
20 3 5 5 24
13 1 3 25
16 26
24 2728
81 0 0 1 1 2 4 2 2 3 3 3 5 3 1 1 2 2 1 4 3 5 6 3 5 6 5 15 Static
97 27 42 46 77 84 98
148
182
228
281
245
336
255
202
207
200
253
280
230
224
235
342
435
169
65 14 0
Inflow to
0 24 16 17 33 44 34 46 64 80 90
134
147
179
225
261
282
339
284
195
219
219
259
351
176
129
178
977 Outflow from
109
10 14 22 20 32 41 45 50 25 46 62 87 77 96
107
105
141
116
85 79
103
116
130
161
195
239
1606 New
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 113
161
195
239
1606 NewUnder30Hrs
34 4 2 5 2 12 12 17 19 7 19 21 24 25 45 56 63
103
74 38 32 45 57 60 37 46 23 32 Transfer
287
37 56 69 98
118
143
195
234
256
330
310
428
335
299
315
307
396
397
319
306
343
464
568
335
266
258
1621 HC2 (Fall 2017)
Table 3. Fall 2016 to Fall 2017 6-SCH M
atrix
33Fall 2019 Volume
Table 3 represents one such matrix, the 6-SCH
matrix from Fall 2016 to Fall 2017. The 28 6-SCH
groupings are labeled down the left with the same
corresponding 28 groupings across the center of
the matrix. This table also contains headcount by
groupings, how many within each grouping did not
return, how many graduated, and how many new
students enrolled in Fall 2017 but not Fall 2016.
Matrices such as this one can be examined to
identify the aforementioned leaks and bulges in the
enrollment pipeline.
The following labels are used in Table 3:
1| HC1 (Fall 2016): Fall 2016 census undergraduate
enrollment excluding special groups.
2| Lost: Enrolled Fall 2016 but not in Fall 2017.
This includes students that graduated without
reenrolling, as a subset. When determining
if the student returned in Fall 2017, only
undergraduate students, excluding special
groups, were considered.
3| Continuing: Enrolled in Fall 2016 and Fall 2017.
4| GradA16: Awarded an associate degree in
Fall, Spring, or Summer of Academic Year
2016–17. Note that only one degree is counted
per student to avoid double-counting, with
bachelor’s degrees given precedence over
associate’s degrees.
5| GradA16E: Awarded an associate’s degree and
enrolled in next Fall term in another degree
program. These students are a subset of
GradA16.
6| GradB16: Awarded a bachelor’s degree in Fall,
Spring, or Summer of Academic Year 2016–17.
7| GradB16E: Awarded a bachelor’s degree and
enrolled in next Fall term in another degree
program. These students are a subset of
GradB16.
8| Columns in the center indicate movement of
continuing students from the Fall 2016 SCH
categories to the Fall 2017 SCH categories. Note
that the central portion of Table 3 does not
include counts for students who enrolled both
semesters but remained in the same SCH level;
these counts are instead separately labeled
Static.
9| Static: Enrolled in Fall 2016 and Fall 2017 and
stayed in the same SCH level.
10| Inflow to: Enrolled in Fall 2016 within a different
SCH level but moved to the current SCH level in
Fall 2017.
11| Outflow from: Enrolled in the SCH level during
Fall 2016 but moved to another SCH level in Fall
2017.
12| New: Enrolled in Fall 2017 but did not enroll in
Fall 2016. (NewUnder30Hrs and Transfer are
subsets of New.)
13| NewUnder30Hrs: New students with fewer than
30 hours.
14| Transfer: Transfer students.
15| HC2: Fall 2017 census undergraduate
enrollment excluding special groups.
According to the table, in Fall 2016 there were 1,589
students in the (0–6) SCH group. Out of these, 597
did not return the next Fall semester. A total of
408 of these students transitioned into the (25–30)
SCH group, indicating that they were progressing
normally, while 232 transitioned into groups of 24
or fewer SCH. With a quick examination of the flow,
it is easy to see that the majority of students are
not returning within the SCH groupings that make
up the freshman and sophomore years as denoted
in the Lost column. In the (85–90) SCH grouping,
109 students graduated, and 5 of the students
who graduated reenrolled in Fall 2017, meaning
34Fall 2019 Volume
that 104 of the students who graduated did not
reenroll. A total of 165 students in the (85–90) SCH
grouping were lost (did not reenroll); subtracting the
aforementioned 104 students leaves 61 students
who neither graduated nor reenrolled.
A total of 914 new transfer students entered for Fall
2017, indicating a significant number of students
who took some type of transfer credit. Many of
these new transfers could constitute dual-enrolled
students who took both high school and college
classes. The bulk of the new transfer students,
however, are entering with more than 54 and fewer
than 84 SCHs.
In observing the higher groupings, the table
indicates that 865 students had accumulated more
than 126 SCH and 448 (52%) graduated. Of the
students who earned more than 126 SCHs, 608 did
not reenroll in the institution.
While this table represents only one of the six
matrices created for this study, the possibilities of
tracking student flow by groupings, classifications,
or years are numerous. Moreover, it can be argued
that the process of tracking student flow through
transitional states within the Markov process is
somewhat intuitive and indicative of the strong
predictive properties of the model.
Table 4 shows the predictions for the next 3 years,
along with the actual data. The model was built using
the flow of students over a particular academic
year. There were six such academic years used for
Table 4. Actual Enrollment and Predictions, Fall 2012–Fall 2017
Model Fall 12 Fall 13 Fall 14 Fall 15 Fall 16 Fall 17
Reality Actual Headcounts 9,475 9,271 9,002 8,915 9,116 9,090
Model 10_11 6SCH Predicted Headcounts % Diff. from Actual
9,948 4.99%
9,999 7.85%
10,002 11.11%
Model 11_12 6SCH Predicted Headcounts % Diff. from Actual
9,244 –0.29%
9,076 0.82%
8,958 0.48%
Model 12_13 6SCH Predicted Headcounts % Diff. from Actual
9,105 1.14%
8,980 0.73%
8,903 –2.34%
Model 13_14 6SCH Predicted Headcounts % Diff. from Actual
8,839 –0.85%
8,745 –4.07%
8,694 –4.36%
Model 14_15 6SCH Predicted Headcounts % Diff. from Actual
8,874 –2.65%
8,865 –2.48%
Model 15_16 6SCH Predicted Headcounts % Diff. from Actual
9,258 1.85%
Note: The model creates predictions for the next 3 years (when actual data are available for comparison) for each of the models
using the 6-SCH methods.
35Fall 2019 Volume
construction of the models. The columns of Table 4
show the years for which an enrollment prediction
was generated. As can be seen in the table,
predictions for the 10_11 Model for both methods
were overspecified by about 5% for Fall 2012
and about 11% for Fall 2014. The 11_12 Models
produced better projections, coming within less than
1% of the actual values for all 3 years. The prediction
of the 12_13 Model differed from the actual
enrollment by an average of –0.2%. Results from the
13_14 Model indicate that the prediction differed
by an average of 3.1%. In most cases, predictions
farther into the future from the years used to train
the models have greater residuals, which is to be
expected in any forecasting problem.
We calculated averages of the absolute values of
the percentage differences between the actual and
predicted values for enrollment using the actual and
predicted enrollment from Table 4. The percentage
difference between the predicted and actual value is
defined as
% difference = predicted value - actual value
actual valuex 100%
We can examine the predictive ability of the models
by using the average value of the absolute values
of these percentage differences, because these
values show on average how far off the models
were, regardless of sign. In a mathematical sense,
the absolute value between two numbers is known
as the standard Euclidean distance between two
points and indicates the real distance between two
numbers (Bartle & Sherbert, 2011). The results as
shown in Table 5 clearly indicate that the predictive
ability of the model decreases as number of
years out from the years used to build the model
increases, which is expected, similar to how weather
forecasts become less accurate the farther they go
into the future.
Based on the results from Table 5, the study will
examine only 1-year-out predictions, because these
were the most accurate. The actual values are
compared with those 1-year-out predictions in Table
6. The predicted enrollment for Fall X in Table 6 is
produced from the enrollment data from Fall X-2
and Fall X-1 and subsequent flow rates from Fall X-2
to Fall X-1.
Note that the 6-year average of the absolute
values of the percentage differences by class range
from 2.8% to 4.7%. The 2016 freshman percent
difference of –12.9% represents an outlier due
to a major university initiative to increase new
freshmen enrollment. This influx of new freshmen
was significantly different from past years and clearly
signals the bulge in the student flow pipeline as
mentioned above. By utilizing the iterative process
of producing Fall X projections from the enrollment
data from Fall X-2 and subsequent flow rates
from Fall X-2 to Fall X-1, the effect of this bulge in
the system can be tracked into the future to plan
upcoming course offerings.
Table 5. Mean Absolute Value of Percent Differences by Years Out for 6-SCH Models
Prediction Time Frame
Mean Absolute Value of Percent Difference
1 year out 1.96%
2 years out 3.19%
3 years out 4.57%
36Fall 2019 Volume
By observing the predictive capabilities of the model,
it is easy to see how administrators and enrollment
managers can use these results to plan for classes
and instructional personnel. Here, both annual
projections and classification average projections for
the 5-year period were off by no more than 6.6%,
which should fall within the margin of error for most
larger institutions.
Furthermore, Monte Carlo simulation could be
used to obtain enrollment predictions that give a
range of plausible values instead of a single point
estimate for a future year’s enrollment. Monte Carlo
simulations have been used in the context of higher
education by Torres, Crichigno, and Sanchez (2018)
to examine degree plans for potential bottlenecks.
In applying these methods to this enrollment model,
the fractions of students transitioning between
specific levels would be treated more like the result
of many coin flips than as fixed fractional values, and
the ranges of predicted values could be obtained by
repeated random simulation. This level of simulation
was not performed in this study.
Table 6. The 6-SCH Models’ 1-Year-Out Predictions Compared to Actual Enrollment, 2012–17
Freshman Sophomore Juniors Seniors All Levels
Mean Absolute % Difference of
Class Levels
2012 Actual Predicted % Difference
2,876 3,114 8.28%
2,035 2,090 2.68%
1,871 1,966 5.10%
2,693 2,778 3.17%
9,475 9,948 5.00% 4.81%
2013 Actual Predicted % Difference
2,729 2,817 3.23%
1,890 1,875
–0.79%
1,870 1,834
–1.92%
2,782 2,718
–2.30%
9,271 9,244
–0.29% 2.06%
2014 Actual Predicted % Difference
2,644 2,709 2.47%
1,803 1,800
–0.16%
1,870 1,789
–4.36%
2,685 2,807 4.53%
9,002 9,105 1.14%
2.88%
2015 Actual Predicted % Difference
2,533 2,574 1.60%
1,944 1,809
–6.93%
1,738 1,816 4.51%
2,700 2,640
–2.24%
8,915 8,839
–0.85% 3.82%
2016 Actual Predicted % Difference
2,921 2,543
–12.93%
1,724 1,885 9.36%
1,855 1,801
–2.89%
2,616 2,644 1.07%
9,116 8,874
–2.65% 6.56%
2017 Actual Predicted % Difference
3,048 3,053 0.17%
1,829 1,788
–2.24%
1,652 1,766 6.91%
2,561 2,651 3.51%
9,090 9,258 1.85% 3.21%
Mean Absolute % Difference
4.78% 3.69% 4.28% 2.80% 1.96%
37Fall 2019 Volume
CONCLUSIONSThe use of Markov chains in projecting enrollment
and the management thereof has gained popularity
among professionals in higher education. The
short-term projections created by this stochastic
process are unique to other time-tested forecasting
tools used in enrollment management. When
used properly, Markov chains can aid institutions
in determining progression of students that
are different from more-traditional ARIMA and
regression prediction tools in that:
1| they can give accurate enrollment predictions
with only 2 previous years’ data, which can be
helpful when large longitudinal databases are
not available;
2| they can be used to generate predictions on
segments of a group of students rather than the
entire population, which may be required for
other models; and
3| the almost intuitive nature of the Markov
chain lends well to changes in student flow
characteristics, which often cannot be explained
by a complex statistical formula.
By creating groupings and tracking students within
those groupings by the state they transition into, the
researcher can also get a better picture of what type
of students are leaving and when they are leaving.
As shown in this study, the strong predictability
of Markov chains allows administrators to better
plan course scheduling and instructor demand
while managing tight budgets. In this study, several
predictive headcount models were developed using
SCH flow as the annual driver. Eight years of Fall
enrollment data from the institution were used to
develop the models. When applied to historical
data each gives 1-year-out predictions within a
calculated level of uncertainty. The models can easily
be modified to change the new student input data,
the continuation rates, and the interlevel flow rates,
should that be desired. Furthermore, similar models
could be used to track Fall to Spring retention as well
as Spring to Fall retention.
REFERENCESBartle, R. G., & Sherbert, D. R. (2011). Introduction
to real analysis. Urbana-Champaign: University of
Illinois.
Borden, V. M. H., & Dalphin J. F. (1998). Simulating
the effect of student profile changes on retention
and graduation rates: A Markov chain analysis. Paper
presented at the Annual Forum of the Association
for Institutional Research, Jacksonville, FL, May 19.
Brezavšček, A., Bach, M., & Baggia, A. (2017). Markov
analysis of students’ performance and academic
progress in higher education. Organizacija, 50(2),
83–95.
Clagett, C. A. (1991). Institutional research: The key
to successful enrollment management. Office of
Institutional Research. Largo, MD: Prince George’s
Community College.
Coomes, M. D. (2000). The historical roots of
enrollment management. In M. D. Coomes (Ed.),
The role student aid plays in enrollment management:
New directions for student services, 89 (5–18). San
Francisco: Jossey-Bass.
Day, J. H. (1997). Enrollment forecasting and revenue
implications for private colleges and universities.
In D. T. Layzell (Ed.), Forecasting and managing
enrollment and revenue: An overview of current trends,
issues, and methods: New directions for institutional
research, 93 (51–65). San Francisco: Jossey-Bass.
38Fall 2019 Volume
Ewell, P. T. (1985). Recruitment, retention and
student flow: A comprehensive approach to
enrollment management. National Center for Higher
Education Management Systems, Monograph 7,
Boulder, CO.
Fallows, J., & Ganeshananthan, V. (2004, October).
The big picture. The Atlantic. https://www.
theatlantic.com/magazine/archive/2004/10/the-big-
picture/303520/
Gagne, L. (2015). Modeling the progress and
retention of international students using Markov
chains. Honors Research Projects 3, University of
Akron, Akron, OH.
Herron, J. (1988). Universities and the myth of
cultural decline. Detroit, MI: Wayne State University
Press.
Hopkins, D. S. P., & Massy, W. F. (1981). Planning
models for higher education. Stanford, CA: Stanford
University Press.
Hossler, D. (1984). Enrollment management: An
integrated approach. New York: College Board.
Johnson, A. L. (2000). The evolution of enrollment
management: A historical perspective. Journal of
College Admission, 166, 4–11.
Liljegren, A., & Saks, M. (Eds.). (2017). Professions and
metaphors: Understanding professions in society. New
York: Routledge, Taylor & Francis Group.
Luna, A. (1999). Using a matrix model for enrollment
management. Planning for Higher Education, 27(3),
19–31.
Oliver, R. M. (1968). Models for predicting gross
enrollments at the University of California. Ford
Foundation Program for Research in University
Administration, Berkeley, CA.
Pierre, C., & Silver, C. (2016). Using a Markov
chain model to understand the behavior of
student retention. In N. Callaos, H. Chu, J. Ferrer,
S. Fernandes, & B. Belkis Sánchez (Eds.), The
7th International Multi-Conference on Complexity,
Informatics, and Cybernetics/The 7th International
Conference on Society and Information Technologies:
Proceedings (pp. 248–251). Winter Garden, FL:
International Institute of Informatics and Systemics.
Seltzer, R. (2017, July 24). State funding cuts matter.
Inside Higher Ed. https://www.insidehighered.com/
news/2017/07/24/new-study-attempts-show-how-
much-state-funding-cuts-push-tuition
Torraco, R. J., & Hamilton, D. W. (2013). The leaking
U.S. educational pipeline and its implications for the
future. Community College Journal of Research and
Practice, 37(3), 237–241.
Torres, D., Crichigno, J., & Sanchez, C. (2018).
Assessing curriculum efficiency through Monte
Carlo simulation. Journal of College Student
Retention: Research, Theory & Practice. https://doi.
org/10.1177/1521025118776618
39Fall 2019 Volume
THANK YOUAIR expresses sincere appreciation for the members who serve as peer reviewers. Authors who submit
materials to AIR Publications receive feedback from AIR members. The following individuals reviewed
manuscripts submitted to this volume of The AIR Professional File.
ABOUT THE AIR PROFESSIONAL FILEThe AIR Professional File features journal-length articles grounded in relevant literature that synthesize current
issues, present new processes or models, or share practical applications related to institutional research. All
submissions are peer-reviewed. For more information about AIR Publications, including The AIR Professional
File and instructions for authors, visit www.airweb.org/collaborate-learn/reports-publications.
Association for Institutional Research
Christine M. Keller, Executive Director and CEO
www.airweb.org
ISSN 2155-7535
Sesime Adanu
David Allen
Rebecca Bell
Timothy Chow
Michael Duggan
Sarah Fitzgerald
Maxwell Kwenda
Ming Li
Mary Milikin
Gordon Mills
Joseph Roy
Steve Simpson
Logan Tice
40Fall 2019 Volume
MORE ABOUT AIRThe Association for Institutional Research (AIR) represents a worldwide community of institutional research,
institutional effectiveness, and other decision-support professionals. Incorporated in 1966, AIR empowers
higher education professionals to effectively and ethically use data for better decisions through education and
professional development, research, and collective action.
A HOLISTIC APPROACH TO IR
This course provides a foundation for participants to meet and navigate the ever-growing demands for data
and information in the current higher education landscape.
IPEDS KEYHOLDER ESSENTIALS: A BEGINNER’S GUIDE
Created for data providers with less than 9 months of experience as keyholders, this course covers basic
concepts and definitions.
IPEDS KEYHOLDER EFFICIENCIES: REDUCING THE REPORTING BURDEN
Created for data providers with 10-24 months of experience, this course is designed to enhance your
knowledge as an IPEDS Keyholder.
WEBINARS
Subject matter experts provide insight into various aspects of IR, IE, Assessment, and related fields.
AIR FORUM
Learn, connect, and share at the largest gathering of IR professionals in the world, and take advantage of
recorded sessions, presentation slides, handouts, and scholarly papers available after the conference.
Association for Institutional Research www.airweb.org